Beyond Mode-Seeking RL: Trajectory-Balance Post-Training for Diffusion Language Models
For practitioners of diffusion language models, TraFL provides a post-training method that avoids the trajectory locking failure mode of reward-maximizing approaches, achieving robust gains across multiple benchmarks.
TraFL introduces a trajectory-balance objective for diffusion language models that prevents over-concentration on narrow denoising paths, enabling consistent improvements over base models across mathematical reasoning and code generation benchmarks, with gains persisting under increased sampling budgets.
Diffusion language models are a promising alternative to autoregressive models, yet post-training methods for them largely adapt reward-maximizing objectives. We identify a central failure mode in this setting we call trajectory locking: sampled reward-driven updates over-concentrate probability mass onto a narrow set of denoising paths, reducing coverage of alternative correct solutions under repeated sampling. To address this, we propose TraFL (Trajectory Flow baLancing), a trajectory-balance objective that trains the policy toward a reward-tilted target distribution anchored to a frozen reference model. We make this practical for diffusion language models with a diffusion-compatible sequence-level surrogate and a learned prompt-dependent normalization. Across mathematical reasoning and code generation benchmarks, TraFL is the only evaluated post-training method that improves over the base model in every benchmark-length setting, with gains that persist as the sampling budget increases. The improvements transfer to held-out evaluations: TraFL stays above the base model on Minerva Math and is the strongest method on every LiveCodeBench difficulty split.